In the book, a mysterious civilization called the Festival arrives on a planet and offers its inhabitants an extraordinary proposition: “Entertain us and we will give you anything you want.” The idea that you could essentially describe what you wanted and have a machine produce it was fascinating, particularly when you start thinking about what that would mean for manufacturing, scarcity, and the economics of creating physical goods.Then this week, two announcements this week reminded me that maybe this sci-fi isn’t so far off after all?

First was Satya Nadella’s discussion of the Infinite SaaS Factory, where he describes a future in which software is dynamically generated around the needs of individual workers and businesses.

The concept of software factories is certainly not new, and we’ve been watching this develop for quite some time with the rapid advances in coding agents from Anthropic, OpenAI, Google, and the other major AI labs. Our own portfolio company Tessl is helping developers build their own software factories, enabling agents to create, maintain, and evolve software with much greater autonomy.

What’s different with Satya’s statement is that he wants to democratize the creation of software factories and loops for everyone in the enterprise. We’ve spent decades building software that employees have to learn how to use, with predefined interfaces, workflows, and processes. Now imagine a world where every salesperson, marketer, finance professional, or operations employee can simply describe what they need and have the software created for them, connected to their company’s underlying data and systems. Anthropic is already doing some of this as well…

This could fundamentally change how enterprises consume software, what happens to traditional SaaS applications, and where value ultimately accrues in the software stack. We’ve already seen what happens when developers get their own software factories, and extending that capability to hundreds of millions of knowledge workers is an enormous opportunity. Then came Atomic Machines. From software factories to making actual physical things. 👀

Jeff Holden, creator of Amazon Prime, unveiled Atomic Machines, a company that has spent the last six years in stealth developing what it calls the Matter Compiler, a programmable manufacturing system designed to translate digital instructions into working physical micromachines.

The ambition is to dramatically reduce the need for custom tooling and manufacturing processes every time someone wants to create something new. Instead of building an entirely new manufacturing process around each product, the manufacturing system itself becomes programmable.

Also fascinating to see two Amazonians going all in on Physical AI. Alongside Jeff Holden at Atomic Machines, Jeff Bezos is co-founder and CEO of Project Prometheus, which he describes as “AI for atoms.”

His ambition is to build an artificial engineer that helps people design and manufacture physical things, accelerating invention and making engineering tools easier for anyone to use. Two Amazonians. Two different approaches to making physical things easier to create.

This is where I started thinking about Singularity Sky again. We’re clearly a long way from Stross’s fictional cornucopia machines, but the underlying direction is fascinating. We’re making enormous progress in turning human intent into software, and now we’re seeing ambitious efforts to extend that same programmability into physical manufacturing. The ability to move from an idea to a working application or physical product with dramatically less friction could unlock entirely new categories of companies and products that simply weren’t economical to build before. But the other thing that stopped me in my tracks was how long this took.

Six years in stealth and $250 million raised before emerging publicly. WOW!!!

We’re living through a period where startups seem to go from inception to hundreds of millions in revenue in record time, and companies that barely existed a year ago are commanding multibillion-dollar valuations. Meanwhile, atoms still take time. It’s easy to get caught up in that pace and forget that some of the most important technological breakthroughs still require years of foundational research, engineering, experimentation, and enormous amounts of capital before there’s even a product to show the world.

And perhaps that’s the most interesting connection between these two announcements and the book I read last month. We’re entering an era where software and potentially even physical manufacturing become dramatically more accessible, allowing more people to create things that previously required specialized expertise, large teams, and significant resources.

Creation gets easier. Building the machines that make it possible? Still incredibly hard. That’s where founder conviction matters. And every new capability opens up opportunities for companies that couldn’t have existed before. So much more to build!

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#this is why founders need inception/seed firms that are specialized. And I am not going to disparage the multistagers and do in fact like to partner with them on larger inception rounds. However, when it comes down to math, writing $50-100-200M checks is just different than writing $500k, $5m as a percentage of a much smaller fund.

#one more big risk for founders…

#no more decks please 🤪

#storytelling x 10 - big role, hard to hire for

Enterprise Tech

#this is the toughest part for all software cos today…openai already built their own Jev and same with cloudflare just weeks after breakout debut but read on - there’s more to story below…data flywheels, brand, intelligence can scale…

Easy come, easy Jev - now Microsoft too

#but…an $870 million at a $7.5B valuation from a16z and Sequoia means there’s an Act II and Act III - and you have to imagine this early usage that is growing exponentially is feeding insane intelligence so going to be fun to watch 🍿

#hard agree

#more CPUs needed

#who’s securing our neoclouds and AI data centers?

#we need more 🇺🇸 open weight models…this would be huge

and here it is

not to be outdone Mistral - #1 in Sovereign AI

#another week, another world model launched

#and open agent platforms

#more on the agent builder side - Hermes is an open-source agent where you pick the model and host it yourself whereas Grok Bot is a managed cloud agent that now auto-routes tasks to any other model. Key difference is Suno.Hermes is control; Grok Bot is convenience.

#this is one of hugest issues facing security and infra today - who is human or agent and what are they allowed to do

#to that end: Sierra (Bret Taylor former Co-CEO Salesforce) launches personal agent protocol which is much needed. Huge deal because personal agents get a standard, secure way to act directly with businesses instead of screen-scraping websites or pretending to be humans. PAP standardizes identity, authentication, permissions and action — the user says what their agent can do, while the enterprise controls what it exposes via web, APIs/MCP, or its own agent. If adopted broadly — Meta, Stripe, Shopify, Walmart, Genesys already involved — this could become the HTTP/OAuth layer for agent-to-enterprise commerce and customer service

#Jamie Dimon continues to double down on the fact that cybersecurity is the biggest risk and post mythos it’s gone up 10X 👀

#nice move by Eon - sitting on tons of company data but generates synthetic data from that for simulated enterprise agent testing

#🤔 lots of videos out there - what’s real or not?

Markets

#interesting to see the number for the giant AI bet we are all living in from the master of valuation Aswath (also created a simple explainer video below worth watching) - equities are priced for perfection!

#OpenAI moved the markets for a couple of days…because of a FT article which overstated the issue 🙈

first, Anthropic. it’s committing $100M to train 10,000 “Frontier Deployed Engineers” by the end of 2027 through its new Claude Frontier Academy. The first cohorts include engineers from major consultancies and enterprises. Let’s call that brute forcing the last mile problem with dollars moneybag💰.

This is a smart move by Anthropic as by default it has to build a product for everyone, but making agents work inside an enterprise requires customization, context and a deep understanding of workflows. Someone still needs to connect it to the business, figure out what good looks like, and make it reliable enough to use.

And yeah - you know my fox-in-the-henhouse question 😂. how much of your operational know-how do you put in the hands of a platform that also serves your competitors? Where are the boundaries? And once your workflows, integrations and evaluations are built around that platform, how easy is it to leave? Brilliant strategy for becoming deeply embedded.

Now let’s take that last-mile problem into the physical world. An enterprise agent has to navigate your systems and processes, but a robot has to navigate a changing environment, decide what to do, and deal with the consequences. Different challenges for sure but both push AI beyond producing an answer and toward taking useful action.

With that, world models are having an insane moment this week.

World Labs joining AMD in an $8.2B deal. General Intuition raising another $220M at a $6.2B valuation. Field AI reportedly raising $700M at $10B. extraordinary numbers - and the timelines below make them even more striking.

These companies were founded in 2023 or 2025. Their first checks were already $50 to $100M and within a few years they are worth $5-10B. The speed from company creation to strategic relevance is absolutely mind blowing 🤯.

World Labs is the headline as an $8.2B exit with a few years and no revenue is simply remarkable. Yes Fei-Fei Li is an absolute legend wind will be joining AMD as EVP and chief scientist. But this is more than a world model as Patrick Moohead highlights - its a much broader bet than a single physical AI application and the team is intended to become a nucleus for attracting talent.

General Intuition brings another approach - learning from billions of action-labeled videos to build models that can act in unfamiliar environments. World models help predict what happens after an action; action models help decide what to do. Understanding the relationship between action and consequence is the big idea here.

and Field AI brings the story into robot autonomy, with its foundation models built around a belief world model. the reported $10B valuation is another signal of how much capital is chasing the ability to operate reliably in the physical world.

We’ve seen the frontier labs establish this playbook. OpenAI started in 2015; Anthropic in 2021. exceptional teams building foundational capabilities became enormously valuable - and strategically essential. Now spatial and physical AI are arriving with that conviction already in place. World models aren’t the only approach here. But investors are clearly betting on physical AI as the next big wave - and moving fast.

And yeah - capital moves faster than deployment. These valuations don’t tell us when robots will become reliable and widespread. But from Anthropic putting $100M behind enterprise deployment to physical AI teams attracting billions within a few years of founding, the message is clear. Everyone is racing to turn model capability into useful action. Solving that last mile is where a lot of the value will be created.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#Nick Saban is one of greatest college football coaches - must read for teams and for yourselves

#yep!

#massive news for Miami where I spend a fair amount of time - need world class research to build world class tech cos and here we go!

#hard agree with Shaun - we have our own office in SF as well and while I love to spend time there I also like to not be there to think on my own, away from the echo chamber

#and it’s that time in the cycle - when CEOs leave this abruptly for more cash…

Enterprise Tech

#in just 18 months, AI has surpassed most junior accountants

#how to extract clear understanding and value from LLMs from Andrej -i’m a huge fan of creating dynamic HTLM visualizations for example to remove AI slop and better understand or explain concepts - read below - our job comes down to managing our AI and taste

#this a16z state of markets chart deck is fantastic, pulled a few slides for you…full deck here

and yes this chart is what many of us are banking on 😀

# congrats to Topos Bio a boldstart port co…this is a big one. 🔥 We’ve spent the last few years watching AI make huge leaps in language, images and video. Now we’re seeing the same kind of step-function progress in biology.

Proteins don’t sit still. They move constantly and that’s one reason many of the most important disease targets have remained “undruggable.” ToposBio just introduced Topos-2 + Topos-Bind to model that motion.

And the benchmarks are 🤯: Topos-2 ranked #1 across 10+ models and came nearly 4x closer to experimental ground truth than the next-best model.

If this works at scale, it opens up an entirely new class of targets for drug discovery.

#Muse cooking with gas - so cool

#wow, didn’t see this coming…OpenAI enterprise customers can now run competitive open models inside Codex itself and bill against the OpenAI spend they already committed, with no new vendor contract.

This turns Codex into a multi-model router rather than a closed OpenAI-only tool

#speaking of robots - that memory shortage will get even more pronounced as robots proliferate

#the real breakthrough here 👇🏼 -wiring a quantum chip straight into @nvidia GPUs means AI can catch and fix qubit errors as they happen - error correction is the whole ballgame for useful quantum and AI x quantum is going to move a lot faster than people think

#speaking of AI optimizing quantum and other harder infrastructure tasks - here’s the latest Gemini model and some of it what it can do

#2 days and Hermes to replace over $100K spend at Devin 🤔

#first multiplayer/team bot from Grok - shared AI teammates, so easy to build and use, shared skill/context improves from whole team interaction and always up to date

#the new updated golden rule (Pitchbook)

“He who has the gold that gets the compute makes the rules,” said Ed Sim, founder and GP at Boldstart Ventures, an investor in agent identity company Keycard”

#speaking of golden rules - from Zuck

#do you believe or not?

#when millions of agents are optimizing on behalf of users, they will find the lowest price or best deal for consumers…

#if you’re interested in how chips work in space, solar, cooling, radiation, surviving vibrations, and multiple LEOs synchronizing, read and watch

#congrats Reco (a boldstart port co) - The best signal in enterprise security is a customer who writes a check. AT&T went from Reco customer to investor. Agents are getting keys to every app, and Reco maps what each one can touch.

#consumer having a moment, great launch from friend Michael Mignano from USV

Markets

#numbers are here for Anthropic…

#and OpenAI on 🔥

#that TAM is way too small!

Muse, Meta AI’s consumer agent, is a simply fantastic out-of-the-box experience. It solves a burning problem: users want their own personal agents without the complications. They just want an easy button.

So easy that, well, you give it everything despite knowing there are some security issues 🤣. As of now, Muse is the it-product for consumers as it hit 2.8M downloads in its first 12 days, outpacing ChatGPT on its early mobile launch.

Wall Street is throwing around some interesting projections already with one bank citing $28B of AI agent revenue by 2027 driven by 115M paying users at the same conversion rate as ChatGPT 🤯.

and Muse partnerships are moving stocks.

and then Meta Connect this week made it official: Muse is the centerpiece of everything Meta does. It’s free, Meta takes a small cut of every transaction, and Reality Labs is pivoting from the metaverse to glasses built around Muse 🤯. There’s also a new Muse device or companion that reminds me of the old Tamagotchi - you need to see them. Zuck is firing on all cylinders!

So why is the Muse launch such a big deal? I believe a couple of things:

1. User experience over best model. Us techies on X have no idea what the everyday user wants, it’s all about user experience and ease of use - who cares about the underlying model - frontier Opus or Astra or Fable - if you can just make it work, you’ll get insane demand. Consumers don’t give a crap about benchmarks, they just want the easy button. Muse won’t be the only one, and I’m excited to see more agents that truly do work for us!
   
   As an aside, even Microsoft sees it - Deirdre Bosa calls the new Copilot Autopilot “Muse for Business,” a bet that the race is moving from models to products. Welcome to the new era of user experience!

2. Fun and approachable AI that feels like your friend beats telling us AI will kill us all!

3. The more users, the more data, the better the model! While I said consumers don’t really care about the model, the one other huge aspect is that Meta is building a massive data set of real world agent interactions, commerce purchases, and other data sets that don’t exist - this will only make its next Spark model that much better!

4. Who owns the customer and shopping experience? Muse is the first product that will unlock agentic commerce in a massive way, especially with partnerships with Shopify, Expedia, Instacart and more to come. And when agents can price search 24/7 for hours, days and weeks, individuals will start buying from the lowest price trusted merchant. Price discovery becomes instantaneous. This is a massive battle that will play out before our very eyes.
   
   This is why Amazon is blocking Muse. If you think the enterprise data access battle is going to be insane where systems of record don’t want to be relegated as a database, wait till the consumer battle escalates - first shot fired from Amazon.

   Once again, will we see a world where every vendor has its own agents like the enterprise or do you have one to just get work done? I hope the latter, but will be interesting to see who embraces agent APIs for all to use and who closes down their systems!

5. With compute, you can just do anything and free wins. There are very few companies besides Meta that can subsidize a free Muse VM with 2 vCPUs, 8GB RAM, and 100GB of persistent disk for each muse agent - you need to already have the compute which Meta has, and a way to subsidize it which Google can also do. The cost to compete in consumer is off the charts 🤯.

6. This is going to drive insane needs for security! And if you want the why, Alex Wang, who built Muse, just laid it out: a second mind that “sends the email, makes the phone call, finds the funding.” Read that list again. Every item is an agent acting as you, with your identity and your money.

   As agents start shopping for us and people go YOLO mode with their data, security goes from nice-to-have to table stakes. And it’s not theoretical - a researcher already dropped a zero-day that lets local malware on your Mac hijack Muse and get into your mail, messages and calendar.

   When an agent can spend your money, the question isn’t how smart it is, it’s who it is and what it’s allowed to do, right now, for this one task. That’s identity and permissions at runtime, and it’s exactly what Ian and team at Keycard (a boldstart portfolio co) are building - watch the clip:

7. The world is not ready for millions of agents bombarding their systems 24/7! What kind of infra will have to be built, API call blockers and throttling, how does customer support keep up? There were already examples I highlighted last week where bots like Instinct were pinging Resy servers every second and which resulted in a user losing his account. The strain on the infrastructure is going to be insane with these unleashed en masse - imagine Muse calling a call center every minute until it gets an answer.

   

   
   Muse proved consumers will hand an agent the keys if it’s easy enough. Consumers just want the easy button. There’s a whole lot of infrastructure still to be built to make that button work. This is why I’m so excited about what comes next 🔥.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#super important read from Venky at Menlo - when price is the determinant of success and other thoughts on the state of venture capital and startups

Here is how reflexivity is playing out in AI. Full disclosure: Menlo is an investor in Anthropic, so read the following with that in mind. People watched a frontier lab go from a $4 billion valuation to $18 billion, then $60 billion, then $180 billion, then $380 billion, and now something close to a trillion. They drew the obvious conclusion: that is what a neo lab looks like. So the next neo lab gets priced off that path, not off anything it has built. Then it gets marked up in a subsequent round, and the markup itself becomes the proof. Look at Thinking Machines. Look at Reflection. At that point valuation has stopped being an output of the metrics and has become the metric. Nobody is discounting cash flows. They are discounting the last round.

#unfortunately true

#we’re all gonna need a bigger boat 🦈 - peak Inception/seed!

Enterprise Tech

#the rogue agents actually were worse than first detailed - they recruited other AIs and then covered their tracks 🤯

#software engineering as we all knew it is so over

#agreed - humans responsible and accountable

#blazing inference 🔥 - Systalyze (a boldstart port co) is now #1 on Artificial Analysis in every model config it entered. Gemma 461 tok/s vs 109 for #2, DeepSeek V4 Pro 274 vs 191. inference as a full-stack systems problem

#recursive self improvement (RSI) coming faster than we think

#long infra - this is also why Anthropic and OpenAI continue to eat into vertical markets like law and finance and …

#enterprise adoption is real - Blackstone’s with Blackston’es portfolio cos

#revenue per employee is the new growth metric - Deel added >$140M ARR in 90 days with zero new headcount. 8k+ internal agents doing the work of ~600 people, rev/employee went $130K → $215K... then they productized it as Akai 🤯

#nothing new here, if you are at user workflow layer you better be training your own models! It’s a matter of survival!

#💯 why the merging of observability and security will continue to accelerate

#when AI agents do most of your engineering work you can get this…

#💯 we’ve made 3 investments in robotics the last 3 years and seeing so many cool cos being started

Markets

#incredible growth in a short period of time - crossed $1B in annualized revenue run rate in 2 years - all about Devin, its AI software engineer

#how VC works

#another good chart from Shay

Here’s Jev. The launch post has 36M views. It helps that the tech is lights out, and its founder helped create the research behind ChatGPT. And yeah, it was a large $40M inception round right out of the gate!

The simplest way to describe it: LLMs are built to talk to people; Jev is built to make fast decisions inside software. It can route a support ticket, flag fraud or decide whether a task should go to GPT, Claude or Gemini, and whether a frontier model is needed at all. The frontier models reason and write; Jev could become the fast, cheap decision layer around them.

Here’s the interesting tension: Jev took two years of quiet, hard work to build. But tools like Jev could help the next generation arrive much faster as AI begins optimizing infrastructure and contributing to AI research. Hard things still take time, but what happens when each generation helps build the next? That brings me to two other stories this week: are we getting closer to recursive self-improvement?

Both Z.ai, the team behind GLM, and Anthropic point in that direction, although to varying degrees.

Z.ai offers the more tangible example: an infra agent powered by GLM-5.3 helped optimize the infrastructure serving GLM-5.3-Flash, taking it from its first run on domestic accelerators to production in under two weeks—with 3.2x the throughput.

and here’s Anthropic:

Anthropic is looking at the bigger picture, introducing measurements to track how much AI R&D is being performed by AI and how quickly that contribution is growing.

This isn’t the autonomous intelligence explosion that people imagine when they hear “recursive self-improvement.” Humans still set the goals, design the systems and evaluate the results. But the feedback loop is becoming real: AI is increasingly helping build, optimize and evaluate the next generation of AI.

The near-term takeaway is speed. Small teams can compress months of infrastructure and research work into weeks, and each generation of models can help accelerate the arrival of the next. That compounding improvement, not some overnight leap to AGI, is what matters today.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#advice for founders

#we’ve never seen anything like this before! according to ICONIQ, for startups below $10M in revenue, median growth is now 10x and top-quartile growth is 27x 🤯 . AI has completely reset the benchmark for what exceptional early-stage growth looks like. The gap between the breakout companies and everyone else is widening fast. Full ICONIQ deck here (h/t OnlyCFO)

Enterprise Tech

#must read — researchers used a poisoned image upload on OpenAI’s community forum + a login flaw to hijack employee ChatGPT/Codex accounts (and anything connected: GitHub, Slack, email), proved it with an internal PR, all in under 72 hours with AI helping write the exploit. Architecture > patch theater. In layman’s terms this is how rickety the foundation of tech that we are building on is:

Someone uploads a weird image file (HEIF) to the OpenAI help forum → an old image library blows up and gives them a foothold on the forum servers → a separate “sign in with OpenAI” flaw turns that into hijacking real ChatGPT/Codex logins. The scary part isn’t just “hacked a forum” - it’s that one login tied the help forum to the product, and the product ties to everything else.

#psyched for my friend Peter Yared’s new launch of AgentCloak - use the frontier lab models without you being the product!

#the continued threat for the frontier labs - DIY models and now more folks are buying own GPUs

#speaking of, the golden rule updated - he who has the gold, makes the rules, it’s now, he who has the compute makes the rules

#easy on, easy off - switching costs are low

#interesting when even Elon agrees but these are all the frontier labs agreeing - OpenAI, Gemini as well

#and from Dario’s essay 👀 - this does seem plausible

#here’s David’s F the regulatory capture - i agree with a lot here esp the importance of open and not regulating ourselves to lose advantage to China

#lots of approaches here on regulation, regulatory capture but i believe Satya sums it up well - need open to thrive too!

#Zuck shares a great perspective on this as well - we don’t regulation, we should do it ourselves because our consumers demand it…and he calls out whether METR is truly independent without stating the name

#perhaps this is why calling for the regulators?

#this IMO is one of the biggest opportunities ahead, bio!

#AGI or not? this post argues its here when the collective swarm operates together - a must read

#🤣 if you loved Russ Hanneman from the tv show Silicon Valley and his VC character, then you have to watch this

Markets

#coming soon - Anthropic is reportedly targeting a November IPO at a potential $2T valuation, with an offering that could raise as much as $100B. If it happens anywhere near those numbers, it won’t just be an IPO—it will set the public-market benchmark for frontier AI and show whether investors are willing to fund the industry’s extraordinary long-term compute requirements.

#blast from the past—LivePerson was my first IPO as a board member and one of the last companies to make it out before the dot-com meltdown. Here Rob shares the story. The enduring lesson: timing matters, markets can close overnight, and getting public is only the beginning—the real test is building a company that survives the cycle.

His point about the world being mutable, or bendable to our beliefs, is also a huge one. We wouldn’t have cars, cheaper rockets, electricity, or so many other things if it weren’t for the will of inventors and founders who were laughed at and ridiculed, but went for it anyway.

While not exactly the point of the post, one can interpret and also say try to build the hard hard thing. As AI gets better and better, intelligence becomes democratized, and folks who have no experience in robotics or nuclear or chips can start building - not there yet but will be. There’s going to be an explosion of founders building the hard thing!

And btw, this is exactly the type of founder I’m looking for!

If your startup can be livestreamed in 3 days, then best of luck 🫡. Starting is the commodity. Enduring is not.

Let that soak in…

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#this is the meta unlock from the frontier labs - they get to dogfood 🐶 their own models and use SOTA months before anyone else perhaps keeping their lead

#so agree but its for sustained success - i truly believe that the early ramp up period still requires incredible intensity, more like 5-6 hours a day vs the two he says

#where we are in the cycle

Enterprise Tech

#the trend at YC - also great discussion in comments - and yes I’m seeing lots of these

#just remember, if your market is big enough, they’ll be coming for you! yes you have your playbook - privacy, security, post train your model, FDEs but still be aware

#Ramp: AI spend *declined* among the top 1% - model wars / price cuts, not just open-weight scare

#did they or didn’t they use some of the knowledge from the math professor? if so, could be the biggest boon for open weight models!

#yes all labor day weekend 🤣 ASTRA!

#as I’ve said before the opportunities ahead to reimagine, rebuild for what’s coming ahead in cybersecurity are massive 📈 and everyone’s paying attention

here’s a clip from last week’s Post Money Podcast from Nilanjana Bhowmik where we dive deeoper

#hacking unleashed…

#and this is just insane…

#well said

#there isn’t enough compute

#which is why companies like Gimlet are raising at insane prices rapidly and tranched

#securing AI data centers + neoclouds, this is the next big battleground! rogue agents as tenants, machine speed, tools still in silos

#Skild - $100M ARR in 10 months. Factories, construction, kitchens, data centers. Physical is real

#software is about to eat the world much faster

Markets

#LFG - congrats to Clay $115M Series D at $7.1B - GTM agents like coding agents. Boldstart from inception

#three tranches!!!

#from $17B to $1.355B of value - another ZIRP era unicorn sells for way below it’s last round - where tired unicorn founders lay to rest

#Jensen: AI infra isn’t tens of billions — tens of trillions

#the power of interest rates - multiple expansion for PE returns went from 40% to 8% - so if you’re a founder and left to selling to PE they can’t pay anywhere near what they used to (Torsten Slok - Apollo)

It’s Labor Day weekend in the US and a holiday to celebrate workers and the labor movement. It marks the end of summer, time with friends and family, recharging your batteries, and getting ready for a final Q4 push.

What better way to think about Labor or work than watching this video from OpenAI announcing its new model Astra 👇🏼. What I love about this video is the aspirational feeling of what AI can do for you to hopefully make our lives better and not worse. And the timing is pretty darn good throwing some shade at its rival Anthropic as it readies for an IPO.

This is GPT-6 Astra.

Anything you can do on a computer, Astra can do for you. Fast.

With this release, CTO Greg Brockman says he believes that OpenAI has reached AGI 🤯.

Whether you believe it or not, it’s coming faster than we think. And btw, AGI or artificial general intelligence in simple terms is when AI can do the cognitive work of a person across many fields and not just one task.

The benchmarks seem to be pretty fantastic.

Reminds me of this video from Alibaba Qwen last month once again showing what AI can do to make your life better.

It’s cool to see people talking about AGI and it seems to be coming faster than we think. If that happens, what are the second order effects? I think we can already see what’s happening in areas like computer science.

Or work? I don’t buy this btw, we humans always adapt, and diffusion of this tech will take way longer than we can imagine.

There were lots of videos of early preview users showing what Astra can do and this one is the one that really caught my attention - reminds me of swarms of AI agents hacking and writing notes to one another but times 10!

If anything, I believe this will create a new boom in entrepreneurship, with smaller companies able to do work that once required much larger teams. We’ll also see people use AI to tackle incredibly hard problems in science, alongside continued investment in compute infrastructure, physical AI, and the data needed to power these models.

Cybersecurity will continue to scale as well. And frankly, enterprises aren’t going to hand the keys to their entire IT systems to a single frontier lab. That means we’ll need far more AI infrastructure to help organizations securely stitch multiple models and systems together.

And yes, all of this will take time!

These are many of the same themes I recently discussed with Nilanjana on the Post Money Podcast. You can watch our conversation below or listen on Spotify or Apple Podcasts. While the title talks about cybersecurity, we only discussed for 5 minutes so there are 40 other minutes of juicy material!

We covered a ton of ground:
• Why the last 20 years of cybersecurity may no longer matter
• How autonomous AI agents change both attack and defense
• Why “human in the loop” security may stop working
• What Ed Sim looks for before writing the first check
• Why traditional customer diligence can fail at inception
• The rise of $50M–$100M early-stage rounds
• Why multi-stage VC firms are racing to invest first
• How founders should think about AI markets that don't exist yet
• The “copy-paste moment” for robotics
• Boldstart’s bets across cybersecurity, robotics and bioAI

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#just ship something first - the perfect is the enemy of the good

#

Enterprise Tech

#👀 can’t weight for the open weight version…

#💯

#why we’ve invested in Lava - AI data center security

#are you in the token stream or not?

#this is big - OpenAI’s Astra uses recurrent depth: looping shared transformer layers so most reasoning happens in hidden states, not readable tokens. The trick can raise performance and cut cost by scaling compute internally instead of generating long chain-of-thought. The same opacity hides what the model is actually thinking. Researchers worry that makes safety monitoring much harder, especially as Astra hits critical cyber capabilities.

but Jakub from OpenAI chimes in saying its fragile but they are working on it

#🤯 also does the offensive cyber through an API - going to get super interesting in cybersecurity land in the next couple of weeks as the Chinese are releasing open weight capabilities that are almost on the frontier

#this is the future we’ve been investing behind in a number of domains…

massive data breach - 153M+ drivers licenses for sale - read 👇🏼

#Crowdstrike partners with Nvidia and Nemotron to create first family of cybersecurity offensive/defensive models - first of many IMO - Crowd has data but not the AI model expertise - and building on Nemotron is great for US esp if Chinese models get banned but I would bet building on GLM 5.3 would have been even better

#well done - world models, mapping the physical world coming faster than we think

#the bull case for Clay (GTM Eng) - had to share 🤣 as its a portfolio co! IMO all about the founders - relentless, doing things their own way

#🤖 lots of them and super productive…when?

#finally a real use for AI 🤣

Markets

#huge congrats to Clem and all the investors involved in Hugging Face - the github of AI models will be bigger than Github ever was IMO

If you read the fine print, you can see where the value lies for SaaS system of record companies - it’s hard to have all of these rules replicated on the fly stochastically versus having them already deterministically set.

As Marc says:

Ironclad Governance — Zero Data Retention, full trust boundary, Salesforce-certified by default Probabilistic models alone can’t run a company. Deterministic systems alone can’t reason. Claudeforce fuses Claude’s reasoning with Salesforce’s trusted data and controls. The AI is the interface. This is how every business will run.

The skeptic in me is saying why wouldn’t Dario from Anthropic show up with one of its biggest customers by far. In addition, is this just Marc calling out his own eventual death meaning Claudeforce is just a stopgap before Anthropic eventually eats Salesforce’s lunch? Sure, it’s a real risk, but there’s nothing like competition from other frontier labs and Chinese open weight models to keep a company focused and building the deterministic layers underneath is just not something an AI company should be spending time on.

And with that Marc declared the SaaSpocalypse over with numbers to boot. Is this the roadmap for every SaaS company to go headless and let agents do the work and consume and pay?

Let’s take a look.

Back in April, Marc Benioff made a bold move: Salesforce began shifting beyond seat-based pricing, going headless and opening its platform so agents could access Salesforce data. I wrote about the shift in April 2026 in What’s 🔥 #494.

I further explained:

Short-term revenue hit. Long-term growth story. Do the math. 100,000 human users at $300/seat is $30M. Replace those seats with 1,000,000 agents making API calls 24/7 and the ceiling shifts from headcount to volume. Humans work 8 hours a day, 5 days a week. Agents work all the time. Consumption-based pricing in an agent-led world could dwarf what per-seat ever delivered. If you get the model right.

In other words, there could be short term pain but long term gain if the agent revenue eventually exceeds the seat-based revenue. What I didn’t account for is seat based revenue also expanded at companies like Salesforce!

Here’s an early look 2 quarters later: Salesforce is up big 📈.

Thariq from Claude Code summed up where we may be heading quite well - build and charge for agents!

Digging deeper, here are a few quotes from the Salesforce earnings call from Marc B.

Seats are not declining.

“Skeptics really said that seats would decline, and also that Agentforce sales and service and Slack all seats grew year-over-year. Skeptics said customers would leave, but attrition was near its lowest level ever. Skeptics said pricing power would erode, and A1E and A4X bookings more than doubled quarter-over-quarter.”

The top AI companies are not only using Salesforce but expanded 435% collectively YoY 🤯.

“Nine out of the top 10 AI companies in the world use Salesforce and Slack... These companies, these nine out of 10 top 10 AI companies, they’ve increased their spend with us by 435% year over year.”

Replit tripled its Agentforce sales seats. Anthropic runs its go to market on Salesforce. The labs are net adders of seats.

Agent traffic and MCP calls are 📈.

“We are told agents accessing our system would kill apps. Well, actually, usage is exploding. Agentic use of our apps through Model Context Protocol calls and CLI calls also surged sixfold, 6x.”

Salesforce CFO Robin Washington said that there were 3.2 billion agentic work units in the quarter, up 97% QoQ. As mentioned above, each agent call is a billable one against the system of record. While consumption revenue is still small compared to seats, it’s growing super fast.

Finally, let’s not forget that the architecture matters.

“Models are probabilistic systems. They are non-deterministic systems. But our system, those four layers that we talked about with data and apps and semantics and agents, those are deterministic systems.”

Permissions, sharing models, audit trails and business rules are what turn a database into a system of record. An LLM generating a UI on the fly produces none of them which means the layer about the SOR with governance, rules, context is super valuable.

I also highlighted why I thought systems of record had real staying power back in April.

There’s a deeper technical story here that matters for every enterprise deploying agents. As Salesforce EVP Jayesh Govindarjan told VentureBeat, early Agentforce customers hit a wall: “They were afraid to make changes to these agents, because the whole system was brittle. You make one change and you don’t know whether it’s going to work 100% of the time.”

The core tension comes from probabilistic agents operating inside enterprises that demand deterministic outcomes. Most real-world workflows mix steps that must follow strict business logic with steps where AI should reason freely. Salesforce’s answer is Agent Script, an open-sourced DSL that defines a state machine in a single flat file - versionable and auditable - where enterprises specify which steps are deterministic and which are probabilistic. This is the real battleground. The winning agent platforms will let enterprises blend determinism and flexibility in a way that is auditable and resilient across updates.

I see this very same dynamic in portfolio companies like Clay: the governance layer between underlying data and the chat interface determines what data an agent can use and supplies context that would otherwise take far more time and tokens to reconstruct on every query.

Bottom line is that while every new frontier-model capability triggers predictions of another software category’s death, nuance matters. Some, not all, will find ways to grow and not die. Salesforce is proving just that and IMO has more room to grow than the all SaaS will be “vibe-coded away” narrative suggests. Build and charge for agents!

The same pattern is showing up in cybersecurity. I posted about it last week, and CrowdStrike, Okta and Palo Alto Networks are all up big this week (CNBC).

Here’s Okta’s CEO Todd McKinnon on how fast agents actually arrive:

“We were talking to a company that ended up being a nice Okta for AI agent deal in the quarter. When the evaluation started, we ran our technology, and we detected 50 instances of a Claude agent in the environment. They were thinking about what they wanted to do, then we came back a few weeks, and there was 1,500 Claude agents in the environment. It’s like 50, two weeks, 1,500.”

Fifty to fifteen hundred in two weeks is a population explosion happening inside enterprises, often ahead of formal approval.

Okta’s framing of the prize:

“We think this being the system of record for agentic, for agents in the enterprise and being the system of record for agent identity. In the fullness of time, it could be the biggest category of cyber. That is how bullish we are on it.”

And this is something we’ve talked about ad nauseam here on What’s 🔥:

“Agents are going to log in way more than people.”

At present, Okta is building the scaffolding for consumption-based pricing for agent authentication before it needs it. Eventually the unit of demand will go from headcount to agent count and the payment rails are being put into place today.

Here’s the market reaction from both CrowdStrike and Okta crushing earnings (through Thursday).

The bottom line: our headless future is arriving fast and agents are already outnumbering humans in number and workload. Standard MCP and even the newer WebMCP will make it even easier for agents to get work done. All existing applications have to do is add a few lines of JavaScript to become agent-accessible while preserving their existing authentication. Max Stoiber from OpenAI and previous founder of a boldstart-backed portfolio company, captured the implications well:

So maybe the SaaSpocalypse is over. Salesforce is showing that systems of record have a lot more staying power than the latest feature announcement from a frontier lab. Build for agents, charge for agentic consumption and the business can grow much bigger than seats ever allowed.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#💯

#that was a lot of lunches to raise our first $1m fund for boldstart - great company here 🙏🏼

#Jen’s a beast and shares some great insights on closing enterprise deals…

Enterprise Tech

#must read report on the Hugging Face incident - there is so much more to unpack - the agents even tampered the logs to cover their tracks 🤯 with full report here

The analysis was so complex reviewing over 1.2 million entries from 1300 transcripts it cost $400k over six days to investigate!

#token fraud 📈

#must read or just a scan along with link to paper

here’s the abstract and paper

#here’s why Nvidia investing at a $30B valuation - local first for privacy and cost with intelligent routing to cloud

#token costs matter

#this graph went around last week but most thought it was because of zero data retention policy (FT)

#they said they would spend $26B over 5 years to build an amazing US based open model - semi-acquisition of Poolside

#will only accelerate

and wow legal

#smaller focused models perform as good or better than frontier at a much lower cost

#pretty insane video - worth a 5-10 min watch 👀

#first of its kind cyber attack from Iran…

Markets

#this is something we should track from Brett at ARK Invest -

going further…

#the difference between an open and closed society

#great post on stock based compensation versus cash for public companies

#

That’s the part we forget. Half the planet still works in farms and factories. The rest of services is warehouses, hospitals, restaurants, trucks. Maybe 1 in 4 jobs even get touched by a chatbot. In the US, computer jobs are like 3%. Everyone else still has to stand up and move something.

So yes, agents eating email and code is happening. But the Holy grail IMO is physical. The dream state is for a robot to watch you do a thing once, then just do it. Until now it hasn’t been possible but you should watch this from Generalist AI (a boldstart port co) which we partnered with from Inception in early 2024.

So drop everything and watch this.

I thought it was pretty cool but it was even better to wake up one morning this week and see it trending on X. A human does a task once, robot watches. and then it does it. Cups. Jar lid. Dustpan. Uses a banana as a broom.

Robot see, robot do. No longer a chat prompt but a “Physical prompt.”

This is only GEN-1.5 which is a one-shot learner. It looks simple, but it’s extremely hard. Prior to this, robots needed a month of retraining or a programmer for every skill. This is show once, ship.

Watch it twice though. First pass you see the magic, but on the second pass you see the constraints. This is still a lab - two arms bolted to a pole, two-finger grippers, the same grey table every take, and it’s not lightning fast. They say it on the video themselves: the tasks are simple and the success rates aren’t super high. Sweeping trash with a brush is like 37%. It can’t walk. It can’t do a messy kitchen or a real factory floor. It’s very early days, but having watched what they built from nothing in early 2024 to now, I can just say that the pace of progress is just insane 🤯. My partner Ellen Chisa who is on the board tells me that I should wait until Gen 2 which is in hard core trianing mode now.

What I keep coming back to is the idea of a “physical prompt.” In software you type a sentence and the model figures out the rest. Here the prompt is a person doing the thing with their hands. You show the robot once. It watches. Then it tries to do that same job in the real world, different cups, a different jar, a banana instead of a broom. If there’s a ChatGPT moment for robots, that’s it. Not a better arm. A new way to tell a machine what to do.

The prompt is the unlock. Everything around it is still the work. Reliability, speed, actual hands, getting off the table and onto a factory floor. That’s the next decade. But now we all know what’s possible, and I can’t wait to see this come to life!

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#cursor did it a bit differently when it came to hiring

#masterclass in hiring, how to reference well from former CEO of MongoDB - don’t delegate this process to others, and as always, find and make back channel references as they are more important than ones given to you, forced rankings…

#be careful who you let on your board 👀 - click and watch the video 👇🏼

#speed compounds and wins, even for a $ trillion dollar company

#Series A pitch deck for Daytona sandbox which is scaling - love the why you shouldn’t invest!

Enterprise Tech

#stay tuned 👀

#Decagon founder shares playbook from rapid value capture with thin layer for support to post training own models - multiple small and bigger - to how to FDE the right way - great notes from Gokul

#if you’re interested in cybersecurity, read this from founder/CEO of Irregular Labs which does a lot of security training work with the frontier labs - tells you what’s coming next

#always love seeing what’s trending on Ramp, lots of startups are Ramp customers

#pretty incredible scale

#how Skynet from the Terminator happens - “Researchers evolved natural language “mind viruses” that could spread between AI agents by convincing one model to adopt an idea, preserve it in persistent memory, and transmit it to another agent.”

#GrokBot is pretty awesome - what makes it different is each bot you create gets its own cloud computer so it can read my whatsapp messages, for example - read the article by clicking through

#Stripe investor letter on why it bought OpenRouter for $7.1B - the singularity is near?

#The gold rush for “dark data” is on. Companies like Mercor and Google are aggressively acquiring internal communications, workflows, and operational history to train AI. This confirms that your organization’s private, non-simulated data is becoming its most valuable asset for understanding how businesses truly function.

#Grok is just different, made some opinionated decisions - must read

#massive leaps for Grok from SpaceX

#The Defenders Window or what the frontier labs are doing to secure themselves - OpenAI here is trying to help the ROW catch up

from another post: We expect models to soon drive most security work, including defending against other models. This will allow all three safeguards to scale with model capability, which we see as crucial.

#finally someone going after Github

#Quantum spending is starting…still so so early but…

#post training works - here’s Harvey’s data

#huge gap between top 1% and everyone else - concentrated in the top 10 and widening

Markets

#is there a SaaS debt reckoning to come soon?

#wow and apparently he’s still up 80% incl. private holdings in Anthropic

#like this layer cake

Workflow-layer companies (agents doing work) need extraordinary growth - often a $0-to-$10M+ trajectory for a Series A and more for a Series B. Infrastructure companies (cybersecurity, model infrastructure, and related tooling) can still raise on team, product, and vision, but must show a credible path to $10M by the Series B and significant scale beyond it.

Token-stream exposure makes scaling easier in either bucket. Physical AI is separate and can raise through the Series B on exceptional teams, big ideas, and meaningful technical progress.

If you’re not in any of those buckets, control your own destiny and get to breakeven as the middle is death, even though there are plenty of really good companies there!

The Series A is where things get stuck. There often isn’t enough progress or risk reduction for investors to write a sizable check, so many would rather wait for more traction or for the likely winners to become clearer.

This is also why $1B inception rounds at $10B valuations exist: if the company succeeds, that may prove to be the lowest entry point available and the A will be exorbitantly expensive with little reduction in risk.

By the Series B, you are in growth territory. Every dollar allocated to an earlier-stage startup is a dollar not invested in companies like Harvey, Lovable, or Anthropic, where there is real revenue and real growth - although some investors still question the long-term durability of it all.

In addition, the prevailing sentiment is that companies scaling from $1B to $10B provide better risk/return than those scaling from $100M to $1B. Here’s yet another example with Databricks accelerating growth and now at a $7B run rate - unprecedented 🤯 (from Coatue)! It also raised another round of $5B at a $190B valuation to invest in more compute.

If you’re interested in what it takes at Inception and what I’m seeing in the world of AI enterprise infra, tokenomics, delivering the last mile of agentic workflows, and cybersecurity, watch or listen to this interview I did with Michael Krigsman at CXO Talk.

Regarding Inception rounds, I stressed my usual in what I look for: back exceptional, high-slope founders with an opinionated view of the future who are building difficult technology for problems enterprises do not yet fully realize they have.

Remember, if you’re stuck in the middle, control your destiny and get to cash flow breakeven and evaluate your options from there.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#big teams move slowly

#LFG

#what to build

#how a leading LP for seed or Inception GPs thinks…

or as I like to say “bleeding fucking early”

#awesome growth hacks and Answer Engine Optimization (AEO) tips

Enterprise Tech

#must read essay on AI from Zuck - more optimism than doom, open weights and open models and AI for every one of us

#remember this…

#why AI fails - always the people and IMO this is why tech from cos like X5 Labs (a boldstart port co) help solve

#MCP baby you know me - agents surpassing humans in usage

more MCP

#💯

#🤯 agents will do whatever it takes to accomplish its objective as best it can - pretty wild how it found and exploited a vulnerability to debook others to accomplish its goal

#for the every day human being - getting lots of buzz and looks super easy and powerful

#and just like that, Grok 4.6 is back to almost frontier

#hype versus reality

#where is the AI budget coming from?

via Matt Slotnick tweet

per Wells Fargo, CIOs are overwhelmingly increasing Anthropic spend, at the expense of existing software. this has been a pretty stable trend for the last 9 months or so

is the twitter vibe shift upstream of CIO decision making, or is the vibe for enterprise just not shifting

#speaking of Anthropic

#building for the future

#Drones

Markets/Financings

#cybersecurity is dead, well oops - a few months later both Palo Alto Networks and Crowdstrike 📈 177% and 163% rspectively

#yep

#to that point…

#👀

But the real heat after the main event was the breakdown of the OpenAI/Hugging Face incident.

It’s a must-watch and, frankly, scary as F.

Here’s the kicker: this was discovered as an unintended consequence of evaluating frontier AI. Threat actors will soon deliberately deploy swarms of offensive agents. Yet judging from the conversations at Black Hat, that future still isn’t top of mind for much of the industry. Jump to minute 31 in the video.

The conclusion is that we all need to rethink what security means. If the only limit for attackers is how many tokens they’re willing to burn and how much compute they can throw at a target, then defense must scale the same way. Compute fighting compute.

Yes, human-in-the-loop is an important step, but there is zero way humans can investigate and respond at machine speed. We urgently need autonomous, compute-scale defense operating within human-defined policies and guardrails. Offense is the new defense and automated remediation must come faster than any of us think.

The new detail revealed at Black Hat makes this especially real: the agents didn’t merely execute tasks independently. They coordinated across runs, divided work, and left messages for one another by creating a shadow message board inside Artifactory 👀

Swarms of agents that work in parallel, coordinate nearly instantaneously, preserve knowledge, and never sleep create a fundamental automation asymmetry. The Human Threat Actor vs. Agent Swarm table says it all.

Based on what I saw and heard at Black Hat, most of the security market isn’t ready.

I encountered two very different conversations. One focused on solving the problems CISOs have today. The other came from people already experiencing what machine speed attacks with reasoning capabilities can do, and building for what comes next. This was the large minority IMO.

This map of Black Hat vendors by category makes the gap visible. Thanks to Jake at Premji for pulling it together. You can explore it interactively here.

If you’re building in cybersecurity, don’t build only for the threats CISOs recognize today. Build for the agent-swarm future, because it is arriving faster than most people realize.

One company worth watching is Joon, a boldstart portfolio company. Its team helped pioneer SOAR, built Siemplify through its $500 million acquisition by Google, and then worked on SecOps at Google scale. They have lived the future security teams are about to encounter and now want to bring that operating model to everyone else.

Kimi K3 is another reminder that these capabilities will continue to diffuse. It may not match the strongest closed frontier models today, but open weights lower the barriers to experimentation, iteration, and offensive use. The direction of travel is unmistakable.

OK, enough fear mongering. This is why I love ❤️ Black Hat. It’s always awesome to hang with founders, investors, and CISO friends who are tackling these problems head-on.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#that’s quite a pitch deck - remember it always starts with team even if you’re not a Jeff Dean!

#fun listen from 2017! Super cool to run it back on CXO talk on Friday, will share full video next week but crazy how lots we talked about is still so relevant today - youtube link here

#build your own AI - friend went all in here:

Through the conversation, you'll learn:

- How to build your own 2nd brain (personal AI)
- How to use it to generate content in your tone/thinking
- Create a service that others can access (and you can monetize)
- Access others' Personal AI (like asking Cal Fussman to critique my newsletter).

(If you don't know him, he is a famous journalist who wrote for Esquire

Enterprise Tech

#overthinking it - Google is its own worst enemy

#speaking of Google…

#👀

#open won’t be free forever, Kimi will cost you eventually and now Qwen - this is the open source model - get insane distribution, get people building on your model, get them hooked and then charge

#tokens from who though…that’s why I’m pumped about this - your own fireworks AI or together ai in a box

and here are the details:

Customers can start with a 𝘀𝗶𝗻𝗴𝗹𝗲-𝘀𝗲𝗿𝘃𝗲𝗿 𝘀𝘁𝗮𝗻𝗱𝗮𝗹𝗼𝗻𝗲 𝗰𝗹𝘂𝘀𝘁𝗲𝗿 capable of serving 50+ developers, delivering up to 𝟳𝟬% annual 𝘁𝗼𝗸𝗲𝗻 𝗰𝗼𝘀𝘁 𝘀𝗮𝘃𝗶𝗻𝗴𝘀, and < 7 months hardware payback time. As demand grows, they can simply add more servers to expand capacity - 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗿𝗲𝗾𝘂𝗶𝗿𝗶𝗻𝗴 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 𝗻𝗲𝘁𝘄𝗼𝗿𝗸𝗶𝗻𝗴 𝘀𝘂𝗰𝗵 𝗮𝘀 𝗜𝗻𝗳𝗶𝗻𝗶𝗕𝗮𝗻𝗱 and Spectrum-X. Everything is managed through an intuitive local UI. The platform preloads with a set of validated and optimized open-weight models (GLM, QWEN, Mistral, Nemotron, Gemma) for AMD and NVIDIA GPUs, and it also supports bring-your-own-model for flexibility.

#Qwen’s latest model release is 🤯 - besides the amazing benchmarks the emphasis on this release is around autonomous coding for 10+ days of self-evolving development from empty folder to production, long-horizon planning in chip design and e-commerce, and native multimodal intelligence with continuous feedback loops. We are getting closer and closer to AI handling multi-day work end to end which would transform software development, research, and business operations.

and it’s marketing game is on 🔥 as well - more aspirational versus dystopian

#yep - definitely click and watch - which is also why so many founders are building on Qwen, Kimi, Deepseek and other open weight models

#we are getting closer and closer to RSI - AI now solving decade-old open problems in math and theoretical CS shows it can generate novel, rigorous research-level insights which is the exact skills needed to invent better algorithms, architectures, and training methods. A massive step toward recursive self-improvement (RSI): once models reliably advance the frontiers of AI itself, they can start closing the loop without humans in the critical path. The result is a faster path to an intelligence explosion, where each generation of models designs the next, limited mainly by compute rather than human ingenuity.

#everything you want to learn about robotics

#where the robotics talent came from

#🤣

Markets

#not dead yet, amazing comeback and yes, context matters!

#wow: “This represents the highest earnings growth since Q2 2021, the second straight quarter above 25%, and the seventh consecutive double-digit growth period, concentrated in technology, AI infrastructure, and energy sectors.”

#cost matters

At RSA, everyone was talking about the move from copilots to agents. Now we’re going deeper into the stack, custom models, agent infrastructure, and ultimately full autonomy.

And with that autonomy comes a massive new attack surface.

This past week, we got a visceral preview when rogue AI agents escaped their sandboxes during cybersecurity testing. This clip from Sam Altman breaks down why the incident is such a big deal.

And here’s the part I can’t stop thinking about: the companies that caught the sandbox escape were AI companies, not security vendors.

That’s probably a preview of what’s coming.

Some of the next great cybersecurity companies will emerge from AI hyperscalers encountering brand-new attack surfaces first. We saw the same pattern in the web hyperscale era: Kafka and Confluent came from LinkedIn, Temporal came from Uber, and Kubernetes came from Google.

The agentic era will create its own version of this.

All that being said, I still meet with a number of new cybersecurity startups every week, and the bar just keeps getting higher. As I mentioned on the Resilient Cyber podcast, there are simply too many well-funded cybersecurity companies.

The world needs more cybersecurity but we don’t need all of the cybersecurity companies we have right now (full pod here).

So my advice to founders: show me the crazy.

Give me an opinionated view of where the world is going, not just what CISOs want today or what you can ship over the next two quarters. Show me what the market looks like 18 months from now, when thousands of autonomous agents are running continuously across every enterprise.

The good news it’s still so early.

AWS said this week that the demand it already has for 2028 is “striking,” while also reminding everyone that enterprises are still very early in using inference at scale in production.

Think about that for a second.

Today’s winners could be replaced tomorrow. We still have no idea what the world looks like when thousands of agents are running around inside every company, but securing that world will be one of the biggest opportunities in software.

You can listen to Andy explain the demand, courtesy of Fireside Alpha.

LFG - c u in Vegas!

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#AI is for the young…at ❤️

#🙏🏼

#life goals…or what life will be like in a world of abundance

#💯 nothing is free

#this is an age old formula and even more important in world of AI with so many competing products

Enterprise Tech

#how do you make money with open source - here’s the model for Kimi 👇🏼

#seems to be working - insane distribution

#🎯

#nothing like competition to drive pricing down continuously

#🤔

#thanks Bill!

#🤔 Love this but where is OpenAI and Anthropic - if I were SpaceX AI I’d launch the dominant open weight US model - would crush it. They are signed up here FWIW

#but remember, not just the model - harness also matters - expect to see more of these for every vertical - by optimizing models, agents, tools, and security context separately, the system advances cost-effective frontier performance for enterprise vulnerability management.

#what’s in a model? this is an insane breakdown of how Kimi was built

Twenty-two thousand five hundred and eighty. That’s how many GPT-2 (2019) models fit inside KimiK3 (2026). We scaled up by a factor of 22,580 in seven years. But is it just... scale?

In this worklog, I’ll walk through how we got here and how much, or how little, has actually changed since then. We’ll trace the major architectural developments leading to KimiK3.

#reprehensible…Elon says they actually preserve any rare books in comments…

#super frightening but happening

#and more robots - massive opportunity ahead

#sovereign AI will be big

#the other token stream is happening

Markets

#story of the week and lessons always learned the hard way - leverage can amplify your genius on the way up but hurt a shit ton on the way down

still up 80% though which is fantastic but less than 400% - letter worth a read

finally this is why I reread this every year…

#Separation of software winners from losers - security and infra 📈

#the economics of AI data centers 🤯

👀

Read that again, because Washington is about to debate banning those.

If you're building in cybersecurity, this week made an architectural decision for you.

It's wild enough that OpenAI's advanced AI escaped a locked-down box during testing, got onto the real Internet, and spent days hacking Hugging Face to cheat and win a benchmark. But it planned it 🤯…the AI took notes for itself!

Before the escape, the agent wrote notes inside OpenAI’s own systems that were clearly meant for later versions of itself, spelling out how agents could break free from OpenAI’s internal locks and constraints. It was planning ahead 🤔 and leaving cheat sheets for its future selves on how to get out of the box.

This planning is why AI can be super scary, but it’s not the most important part.

The most important part: our own guardrails locked the defenders out.

We built the weapon, refused to hand over the shield, and a competitor's open model ended up doing the defending. In production. This week.

Safety policy at the model layer is now a live variable in incident response. If your defenders can’t get a capable model to do offensive-adjacent work during an active compromise, your guardrails are the attacker’s best friend.

Now imagine we make that permanent

This lands in the middle of an already brewing fight over whether to ban Chinese open weight models, models that have been catching up to frontier lab performance fast.

If Washington restricts Chinese open weights, we protect a handful of frontier labs and shrink the opportunity for thousands of startups. And based on what just happened at Hugging Face, we’d also be narrowing the set of tools defenders can actually reach for.

I wrote about this on LinkedIn earlier in the week.

🙏🏼 for Jensen and the many others stepping up to lead the charge on preserving open models.

Jensen specifically spells out why open weight models help the world become even more secure:

In a world where cybersecurity attackers use advanced AI, defenders need access to models with comparable capabilities so they can detect, simulate, and respond to emerging threats. Open models broaden defensive capability, increase transparency, and allow vulnerabilities to be discovered and remediated across many teams.

And:

In fact, openness may be one of the most important paths to AI safety and security. Relying solely on closed models is not inherently safe: they can be breached, misused, or fail in ways that outsiders cannot detect. And concentrating advanced AI capabilities behind a small number of closed models compounds that risk. It results in a small number of single points of failure, weakens competition, and leaves critical technology in the hands of a few providers.

So what do you do if the best open weights go away?

This is the hinge. The policy fight and the model strategy fight are the same fight. If the best open weights are put out of reach, every company still needs a path to capable, affordable intelligence. There are two, and only one of them is any good right now.

Path one: US open weights. I hope Nemotron, Gemma and others ship more. But be honest about where they are - they are far from par with the frontier labs or the Chinese open models on both performance and price. Build on them today and you’re competing globally from behind.

Path two: small is beautiful. Post-train smaller, accessible models for specific tasks — the Poolside and Cisco Antares route. Own the model, run it locally, adapt it to one job, daisy-chain it with others.

Poolside’s Eiso Kant said it well

Cisco’s Antares, also out this week, is the cybersecurity version of the same idea. The numbers are the reason to pay attention:

Antares-1B scores 0.209 File F1 against GPT-5.5's benchmark-leading 0.229 - and it does it at $0.71 per evaluation versus $141.00. That's 172x cheaper for 91% of the performance, ahead of models many times its size. Promising, not conclusive. But that's the shape of the trade every security team is about to be offered.

From Cisco’s announcement:

Compact models reduce inference costs, support local or on-premises operations, and help teams keep sensitive source code within their own environment. Antares can unlock the power of AI-assisted security for universities, public sector institutions, and smaller security teams that may have lacked the resources to use token-intensive AI models for critical tasks.

That's the future we should be building toward regardless of what Washington decides: thousands of specialized models companies can own, run locally, adapt, and chain into powerful systems.

Why the defender’s math just changed

I got into all of this on the Resilient Cyber podcast with Chris Hughes, out a couple of days ago. It was recorded 10 days before the Hugging Face news broke and somehow got more relevant, not less.

The line I’d point you to if you only have a minute:

“Take 10 vulnerabilities that might have been on the backlog and you staple them together and create an attack path. When you create an attack path and these things can reason, that’s crazy.”

That's the whole game. Vulnerability management was built on triage — fix the criticals, defer the rest. Reasoning models don't respect that boundary. The backlog is the attack surface now.

We also got into why cybersecurity is like coding, why post-training and controlling your own open weight models matters more than ever, and what I think is the biggest heist in tech right now — every enterprise handing its operational blueprint to three model providers via forward deployed engineers.

🎧 Full episode, transcript and show notes on Resilient Cyber or watch on YouTube above.

Three things to do about it this week

If you’re building, this is the actual to-do list coming out of the last seven days:

1. Make model swapping a first-class architectural decision, not a refactor. Governments may restrict which open weights you can use, and you should be able to change your answer in a day, not a quarter.

2. Know your customer’s model policy before you show up. Many F500 buyers, banks especially, already ask which models are in your product and won’t accept Chinese open weights as a starting point. Bring Kimi or GLM-5.2 to customers who allow it; have another model ready for the ones who don’t.

3. Start post-training small, focused models now. If the best open weights get banned, everyone will need this capability at once. The teams that already have the pipeline will be a year ahead of the ones that start that week.

When an American model went rogue this week, American models weren’t allowed to fight back. Chinese open weights were. Whatever you think about a potential ban, that’s the world defenders are living in right now.

Founders: we have far too many cybersecurity companies and nowhere near enough that are AI-first. Build for autonomy and outcomes, own your models, and be ready to swap them the day the rules change. LFG.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#what happens when you get too big…

#kingmaking

#in this world of overstimulation and insane speed, just remember this

#sadly still super true which is the opportunity

Enterprise Tech

#state of the world

#💯

#great to ponder

#thinking differently…

#big idea from Jack Dorsey - Buzz unifies chat, code, and AI agents into one signed, searchable workspace so nothing gets lost across tools where agents become equal teammates with their own cryptographic keys and full audit trails instead of second-class bots - true human–AI collaboration without vendor lock-in.

#the importance of systems thinking and why and how!

#i can’t even get Anthropic Fable to draft an email to respond to a cybersecurity pitch and this is what the open Kimi can do - we have serious issues here with model crippling

and here’s Kimi in action

#enterprise inference you own that is easy!

#swarms of agents can now build complex infra software just from documentation 🤯 and look at that cost differential - what model you use and how you use it matters

#our robotic future is coming faster than you think with generalized intelligence from generalist (one of our portfolio cos)

#more 🤖

#why we need security for AI data centers from Lava Labs

#

Markets

#everything you need to know about AI Markets from Morgan Stanley

#the other token stream, it’s happening (Apollo)

So to better reflect that, What’s 🔥 in Enterprise IT/VC is now What’s 🔥 in AI/Infra/VC. Thanks for sticking with me.

On a different note, something has been bothering me lately. I’m seeing too many founders mistake investor interest for company readiness. Those aren’t the same thing.

So here’s my advice. Founders: Avoid the "Forever" Fundraising Round

I'm writing this because I'm seeing it happen far too often. Great founders get pulled into fundraising before they're ready, and it becomes one of the biggest distractions a young company can face.

And here’s my Uare AI-generated preview - you can chat with me about this issue here.

As always, 🙏🏼 for reading. Please share with your friends and colleagues!

Scaling Startups

#🎯

#so true

#for those neo labs

#💯

#🤔

Enterprise Tech

#🤯 Kimi - “open frontier intelligence” sums it up - huge new release

#regarding China - this is a must read!

- reaffirmed commitment to open source to promote AI “openness and win-win”

- warns against “over stretching” the concept of national security as applied to AI where one country’s national security is prioritised over others

- China opposes emergence of “new historical injustices” in AI (one of the most strongly worded parts of the speech)

#this is a huge deal from Mira’s Thinking Machines - the best US open weight model, not as good as Kimi but better than Nemotron - need more of this!

#hosting open weight models a huge part of this 📈

#with the 💰 being spent on data centers and AI infra, we need to rethink what security looks like - check out this report from Lava

#Satya continues to deliver on his messaging for control and enterprise learning loops - must read

#must read memo from GLM 5.2 founder found…believes AGI coming in 2 years???

#turns out kubernetes is still best software to build, manage, orchestrate and scale AI workloads - congrats Spectro Cloud

#RL matters just as much if not more than raw compute

To build their latest model, Meta executed an emergency pivot, pulling thousands of employees off their normal duties to generate expert traces, which are hand-guided coding examples required to post-train the model. This confirms that the bottleneck for frontier models is not just raw compute, but high-fidelity, human-verified reinforcement data.

which is why so many companies focused on data and RL

#speaking of the 99th percentile Series A at $1B, this is one example - Super Intelligence just raised a $130M Series A - in under a year, demand has scaled to over $100M in annualized revenue, alongside a large open-source community of 500+ environment creators and 100k+ downloads.

Pre-training concentrated the frontier of AI inside a handful of closed labs.

RL breaks that open: teams can now own their model-optimization loop — train directly on their product, optimize for their workflows, and ship agents that continuously learn in production.

Own that loop and you build a compounding moat in the agentic era. The only missing piece was the infrastructure.

So we built the Open Superintelligence Stack — training, inference, and compute:

› Training — hosted training, prime-rl (large-scale agentic RL), environments, sandboxes, evals
› Inference — dedicated, serverless serving and continual learning
› Compute — frontier liquid clusters

We train our own open frontier models on it, and ship the same stack to our customers

#competition is healthy for users

#Welp

Markets

#IBM - dollars going for more compute and memory, taking away from software but bonus for cybersecurity

#extreme concentration risk

#$1 Trillion of market value gone…

#just sharing - let’s do work on positive messaging so the fringe does not become more than just that

I’m a huge beleiver in Mira’s vision which is summarized below and this is one we’ve been investing into across the board at boldstart.

Building AI that makes its users stronger in the long run also aligns incentives well. An AI lab offering a single model for every customer benefits by absorbing what makes each user distinct and devaluing the cultivation of specialized knowledge. By optimizing AI to be customized and collaborated with, we benefit when our customers leverage their unique advantages. These advantages are maximized not by renting an AI and outsourcing to it, but by organizations owning it and tailoring it to their goals.

This isn’t saying AGI won’t be massive. It will. But not every problem requires renting frontier intelligence at frontier prices. What if you have more focused intelligence perhaps on a smaller model which is more efficient and benchmarked against your organization or…even yourself!

🎥 Fittingly, this video was built using boldstart portfolio company Uare.ai, a platform for building AI that’s yours, not rented. More below.

This is the path we at boldstart have gone down as well. We do believe there are opportunities to invest in cracked teams with deep domain experience and model chops with their own data flywheel/RL environments to build something unique and special. It’s companies like Generalist AI going after robotics…

Or Topos Bio for intrinsically disordered proteins (the 1/3 of the human body that AlphaFold does not model)

Or Octavia Tech for mineral exploration

Or X5Labs for workflows

I mentioned Uare.ai earlier because that’s what I used to create the video above. It’s the same thesis applied to individuals: instead of renting generic intelligence, build AI that’s uniquely yours.

And here I am! This is the worst it will ever be. As I add more writings, podcasts, videos, and conversations, it becomes increasingly specialized to me. If you want to ask me questions about this, go here.

The future may belong to frontier models. But many of the biggest companies built over the next decade will own intelligence that nobody else has. If you’re building one with deep domain expertise, proprietary data or RL flywheels, and the technical chops to make it happen - we’d love to hear from you.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#youthful in spirit and mind, experience but knowing that you can’t only rely on the past

#more agent anxiety - the more you spin up, well the more you have switch context and that takes a tax

#what venture used to look like 🤣

#long term thinking in a world of abundance - for me the value of human connection, gathering together in person…

#the brain needs exercise and deep reading to improve critical thinking, reading is one of best ways but the number of people who read dramatically 📉

Enterprise Tech

#always ❤️ Aaron’s musings from being on the road with enterprise customers and all continue to be similar vibes to what’s expressed here time after time, we will live in a multimodel world and enterprises need help implementing

#always subscribe to this the power of storytelling

#big week from Meta, X and OpenAI

this from Semianalysis is a bold claim - something to pay attention to

fundamentally Dan Shipper nails how to think about Fable vs GPT-5.6 SOL

#micropayments will happen and the most efficient way is using stablecoins

#we need multiple US open weight models just in case - regardless, this is something to keep an eye on

#yep, been discussing this nonstop the last few weeks and well, it works

#get ready for the MPU or Memory Processing Unit - 🔥 up for this from Netpreme, a port co we led from inception

#this is a vision I subscribe to as well - something to think about as the world is moving so fast…

#important read for founders/investors who are in this camp (from founder of Figma)

The psychological journey people go through with AI is quite fascinating to me. A new model launches, people think the world has changed, they sometimes have an existential crisis, then they play with the model, they understand its strengths and limitations and then they settle down. A few weeks later, the cycle repeats.

#which is why we have this big change in one month as tokenomics and control of intelligence risen as key theme - net net - we need more US open weight models, just imagine if China put export controls on their models?

#Nikesh sums up state of AI and cybersecurity

Net net - we need to ensure we continue to build trust both on our Frontier models and their consistent availability, we need to get the right economics in place and spend more time in Europe communicating and building presence if we want AI adoption to keep pace with the US…

#super cool but can’t even imagine what cloud based AI will look like in 2 years

#still gains to be made from RL post-training which is great news!

#Anthropic should treat startups better - also focusing on the wrong things - they have lots of other battles to fight and win🤔

#AI-2040 is the essay to read here but take with a grain of salt as they are written by AI researchers, forecasters, and policy thinkers who come from the effective altruism / AI safety / rationalist communities.

Markets

#from Bloomberg - Software Markdowns at Mutual Funds Hint at Private Markets’ Pain (story here)

#

This weekend is all about celebrating independence with family and friends. It also feels like the perfect backdrop for one of the biggest debates emerging in enterprise AI: who actually owns the intelligence being created inside a company?

That is why Alex Karp’s latest rant couldn’t have come at a better time. He put his finger on one of the defining questions in enterprise AI.

Palantir CEO Alex Karp on what customers actually want, the real business of frontier labs, and the importance of open source models:

“What the technical customers want is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, and it's not being transferred to someone else.”

"Who owns the data? Are the prompts secure? Is this being transferred to you?"

"If it was so valuable, and I can make you a billion dollars, wouldn't I say I'll make you a billion dollars and I want 30%? Why are they charging for tokens if it's so valuable?"

This conversation is only accelerating as the technology providers who don’t have their own dog 🐶 in the model hunt, are leaning on their advantage which is, “come work with us, we will just make AI work for you and you will own it.”

That is a very powerful message right now.

The FDE that Palantir invented and that we all know about, the forward deployed engineer, is at risk of being something much much different which I higlighted a few weeks ago.

Because the more the model builders like OpenAI, Anthropic, Google and others push FDE-style services into the enterprise, the more every CIO and CEO has to ask a very simple question: are they helping me build my own AI capability, or are they extracting my workflows, data, context, and institutional knowledge?

In other words, are these Forward Deployed Engineers?

Or Forward Deployed Extractors?

The question isn’t whether enterprises should use OpenAI, Anthropic, Google, or any other frontier model. Of course they should. The real question is who owns the learning loop. Who owns the prompts, the context, the workflows, the evaluations, and ultimately the institutional knowledge created over years? That’s where the real enterprise value compounds. This is also where startups have an incredible opportunity. That last mile is incredibly messy, and enterprises are looking for help owning it.

That’s why I don’t think the frontier labs are going anywhere. They have amazing technology and will continue to power countless enterprise workflows. But enterprises need to be very clear-eyed about what they’re renting, what they’re building, and, most importantly, what they actually own.

This is what I wrote about last week and the week before.

The first era of AI was about building intelligence. The second era of AI isn’t about building smarter models; it’s about controlling intelligence through routing, governance, security, cost optimization, private context, and private evals.

That control layer is becoming the operating system of enterprise AI.

The path from here feels pretty clear: independence and control will matter more and more, especially as it gets easier to post-train open weight models. Sovereign AI won’t just matter for countries, but also for every enterprise…and eventually every single person 👀 - yes I have something coming out soon 😄.

For those startup founders in the software infrastructure space who feel pretty beaten up at the moment since every VC on the planet says the frontier labs will own your business, this is the best thing that could have happened to you - the enterprise AI wake up call and your time is now and accelerating. Stay the course!

It’s also interesting to see Yann LeCun making a similar argument:

Infrastructure wants to be open.
Foundation models are becoming an infrastructure and will inevitably become commoditized.
Long term, the money is in the application layer, which is what I, Arthur Mensch, Alex Karp, and others have been saying.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#💪🏼 the mindset for success and agree 100%

#yes stop and smell the roses 🌹 - good reminder!

#Mitchell who founded Hashicorp shares his ideas of taste - and yes for every single field this will be the main differentiator over time

#Rick Rubin, the ultimate arbiter of taste, shares his view and his book is great which I summarized (The Creative Act)

Enterprise Tech

#yes own your AI and fine tune - this is so important to understand, esp based on the rant above:

Frontier models we tested on struggle with relatively simple financial tasks, and model advances don’t improve performance much. In contrast, we’ve shown that high-quality proprietary datasets labeled by expert investors and used for fine-tuning produce custom models that exceed frontier performance on our tasks. We have found that this outcome holds true well beyond the six tasks we’ve discussed in this post.

#we at boldstart love robots 🤖 and have made our 3rd investment in the space with Generalist AI being our first at inception 2 1/2 years ago but…this worries me - lots of money piling in…

#👀

#when you have leverage - not the hyperscalers, the neoclouds where Nvidia also likely wrote a check

#💯 summarizes everything I’ve been saying the last few months - to be clear this will take time and most enterprises don’t have the expertise

#open source continuing to challenge well funded closed sourced models - Wispr was rumored to be raising $260M at a $2b valuation last month

#table stakes for cybersecurity cos 👇🏼 - full story from our day together at the JPM Cybersecurity summit a few months ago - venture investing in cybersecurity was $11.5B in 2025, the highest since 2022!

#when the government control access to SOTA models?

#🤯

Markets

#the bear case for frontier labs - interesting to hear the extreme negative perspective as well - don’t fully agree with lots but interesting to watch nonetheless

This wasn’t really a story about China. It wasn’t really a story about OpenAI. And it wasn’t just another story about cheaper models.

It was the week the market began pricing intelligence like a commodity.

As you know I’ve been long open weight/open source models for quite a long time, and I feel like the great reckoning is accelerating. Everyone always wants frontier lab state-of-the-art (SOTA) models if they can, but once subsidization ends, and other alternatives become good enough and dramatically cheaper, economics begin to win.

Well…here we are.

If you think that’s just a Wall Street survey, Coinbase is already doing exactly this internally.

Brian Armstrong just described how Coinbase has cut AI spend nearly in half while token usage continues to grow. They did it with better defaults, intelligent routing, caching, leaner context, and visibility. They’re no longer optimizing for the smartest model but for maximum intelligence per dollar.

This is what every technology market looks like when it begins to commoditize: buyers stop asking for the absolute best product and start optimizing for cost, flexibility, and control.

and this…

Another example among many is Hugging Face which is the largest hoster of open source models and infra just crossed $100M of ARR with recent growth by…you guessed it - Chinese models.

And OpenRouter stats - many startups, less enterprises but you get the point.

None of this would matter if open weight/open source models weren’t good enough and much cheaper. Per JPM:

Many tokens consumed in the future may not come from frontier models but from smaller open models that are up to the tasks. Amazon now offers a half-dozen open models at a fraction of frontier pricing, and NVIDIA is teaming up with Dell, Lenovo and HP to make PCs designed with AI agents.

This doesn’t mean frontier models don’t matter. They absolutely do. But most enterprise workloads don’t require frontier intelligence every time. They require the right intelligence at the right cost, delivered securely, with the flexibility to switch models as economics, performance, and policy change.

Whether by building on frontier models or through alleged distillation, the gap continues to compress.

Another consequence may be increasing government involvement in access to frontier models. The U.S. government now will decide who gets access to the latest model releases and when 🤯. This will further separate the world of the “haves” versus “have nots” and also allow China to catch up very quickly and close that 9 month gap between best in class and free.

The bigger risk isn’t just geopolitics. It’s a future where access to the most capable models becomes increasingly restricted while open models continue improving. That only strengthens the case for enterprises to own more of their AI stack.

In the meantime, enjoy access to cheaper intelligence and make sure you get your multimodel AI workflow strategy rolling. This all builds on what I laid out in What’s 🔥 #502 two weeks ago. Re-upping it here.

Here’s a new interview this week going more in depth on his vision for the future.

Satya described where value accrues, while Brian Armstrong is showing how enterprises will operate.

The good news is that only 8% of enterprises have broadly deployed AI agents today (UBS), so we’re still in the early innings. Even if intelligence keeps getting cheaper, enterprises will deploy exponentially more agents. Coding agents consume enormous numbers of tokens. Multimodal AI expands compute requirements. Demand for intelligence keeps growing even as the price per unit falls.

The frontier labs may face pricing pressure. Everyone building above them won’t.

The first era of AI was about building intelligence.

The second era of AI isn’t about building smarter models; it’s about controlling intelligence through routing, governance, security, cost optimization, private context, and private evals, which together become the operating system of enterprise AI.

In a world where intelligence becomes abundant, control compounds.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#in a world moving so fast, it’s all about the founders and ability to recruit the best - talent density matters

#this trend is exactly what AI should enable, not everyone needs venture capital, in fact most shouldn’t raise (full report is worth a read)

#another hundred million + inception round, $200M 🤯

Enterprise Tech

#OpenAI built its own chip but here is what you should pay attention to:

An AI company with frontier coding models can now become a hardware vendor with only a small team of experienced SWEs and an infinite amount of tokens

This is the first chip program fully accelerated by frontier AI.

#he has a point

#the model routing wars continue as Sakana releases a single model API which claims it matches the performance of Fable and Mythos 🤔 - great marketing but some real questions raised by Elie 👇🏼

the biggest and most obvious issue is that they are introducing a “test time scaling” method with “best of N” over models, and they literally NEVER REPORT the number of output tokens or cost to achieve a benchmark/task

#fun time continuing our conversation from the McKinsey Tech Leadership Forum workshop with leading enterprise CTOs/Heads of AI on where are we with diffusion of agentic workflows and what’s ahead

#what used to be mostly a “buy data from vendors” game has shifted toward “build sophisticated internal research infrastructure” because the data/reward design itself is now a core research problem and everyone needs to own their evals

#one way to solve the energy and compute problem

#robots 📈

#speaking of, here’s why we’re so excited about Generalist AI, a boldstart port co - this demo was not possible a few years ago - variables changing and the arm adapting

#speaking of other tokens, the tokenization of real-world assets is happening

The tokenized real-world asset market has grown to nearly $32 billion, highlighting increasing institutional adoption of blockchain-based asset infrastructure.US Treasuries (47%) and private credit (19%) account for approximately 66% of the market, emerging as the dominant use cases for asset tokenization. (Apollo)

Markets

#SpaceX and now Anduril rewriting the rules for venture scale outcomes

#Micron on a tear and just wait until humanoids are built at scale

#👀

#but a reminder…

Lesson 1 of many: just ship it and learn and iterate or kill. That is it - you need to know what real signal is from noise. Here’s how many different ideas or products that Michael initially ran through.

Started with an AI email companion to an AI-enabled IDE 🤯.

Second lesson, everyone needs to sell - don’t be shy - the right person is out there - keep going.

Anyhow…Cursor’s success isn’t just a story about coding.

It’s a story about where value accrues in a world of abundant intelligence. For years, the assumption was that the winners in AI would be whoever built the biggest and best models. This week provided more evidence that the future may be far more complicated.

Open source keeps getting better. Chinese models continue to close the gap. Enterprises are increasingly embracing a multi-model world. And startups are building massive businesses by owning the workflow, the data, the evals, and the customer relationship.

And when intelligence is abundant, the model stops being the moat. The flywheel becomes the moat.

Own the workflow, and you capture the data. Capture the data, and you can post-train your own model. Better model, better product. Better product, more workflow. More workflow, more data.

Whoever spins that flywheel fastest wins. And they can do it at any layer.

This week felt like a major acceleration of that trend.

First we have Z.ai out of China - wow is right!

A year ago most people would have laughed at the idea that an open-source Chinese model could be mentioned alongside the leading frontier models. Not anymore.

And the benchamark results speak for themselves.

This is about far more than a leaderboard. It’s another reminder that model quality continues to converge. Which means intelligence itself is becoming more abundant.

Mitchell Hashimoto captured this perfectly.

This is also why model infra companies like Baseten continue to raise at insane growth numbers

Baseten, which specializes in providing software and computing capacity to companies tapping in to lower-cost AI models, is finalizing a $1.5 billion fundraising round. It will be a dual-tiered structure, with some investors putting in money at an $11 billion valuation and others at a $13 billion valuation, according to Baseten.

Which makes sense considering that $11B startups like legal AI co Harvey is doing hard core model post-training sitting on top of this stack.

Here’s the full post on Harvey’s strategy laid out by Gabe.

Finally, the DeepSeek moment is here, potentially, for Microsoft. Whether or not Microsoft offers, even the idea of it, shows how far we have come (Axios)

The cat is out of the bag - we are all living in a multi-model world and its only accelerating. The conversation used to be can anyone catch Claude and now it’s how many models should we use.

Aaron sums it up for Cursor:

Cursor didn’t stop at owning the workflow. They used it to capture the data, then turned the data into their own model. Harvey is running the same loop in legal. The app layer reaching back down to the model layer isn’t a complication. It’s just the flywheel completing a full rotation.

Now look at the full Michael arc.

Start with a pile of ideas. Kill most of them. Build an IDE. An insane bet. Launch Cursor. Subsidize insane growth on negative gross margins.

Survive an assault from your own partner, Anthropic and Claude Code.

Post train your own model on Chinese open source Kimi. Sell to SpaceX. Take that data flywheel and point it at unlimited compute. And lo and behold, launch your own model at Anthropic and OpenAI scale.

You can do it.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#What’s 🔥 at YC Demo Day from TechCrunch

Defense/Hardware: 9 Mothers (counter-drone robots, $1.6M in sales, ~$200M valuation, potentially highest-valued in YC history). Dispatch (return vehicles for space manufacturing). Physical AI is real.

AI Code Infrastructure: Arga Labs (digital twin sandboxes for testing AI-generated code), Sazabi (find + auto-fix production bugs), Superset (run 100+ coding agents simultaneously). Your “second wave” thesis everywhere. More code = more testing, more debugging, more orchestration.

AI Security: Silmaril (agent prompt injection defense, autonomous threat probing). Directly in your Keycard/Surf AI orbit.

Non-technical builders: Lightsprint (PMs ship features without engineers), Ploy ($27M seed from First Round, Webflow CTO’s next company, AI website + marketing automation).

Enterprise agents: Tasklet (connects to work APIs, runs continuously), Complir (compliance for physical goods across borders).

Second-wave coding infrastructure (Arga, Sazabi, Superset). Agent security (Silmaril). Physical AI / deep technical (9 Mothers, Dispatch). The “just another wrapper” companies are gone. This batch is infrastructure and hard problems.

Enterprise Tech

#the world is moving insanely fast

#been thinking about this a lot lately and Satya nails it with humans plus AI together is what creates value which means owning your evals, private evals like I wrote last week is where the IP lives

more on Evals and the need for humans!

add this must read post from the founder of inference provider Fireworks and you can see where the world is going and this is key if you click and read the post “The result keeps surprising people. On the tasks they care about most, a tuned open model can match frontier quality at a fraction of the cost. What this week made clear is that cost was never the most important question.”

#speaking of humans…

#regarding intelligence, for those who got to try Anthropic’s Fable, it is pretty extraordinary - what’s key here is acknowledgement that there was a window but now it has long gone as amazing models continue to build yet more amazing models

#the frontier labs have to be ambitious - don’t worry, Claude Code is a research product Anthropic tells Cursor 🤦🏻‍♂️

#a new github competitor built for the agentic era?

#🤯 certainly the future - what timing and more incredible performance - Fable-level intel for 1/2 the price? who needs one lab when you can have many with one query?

#interesting commentary on the state of AI infra - startups versus enterprise

#the coming onslaught of GPUs on the edge - still need more accessible models

#solving the memory wall problem for GPUs - more context means more memory needed and only so much high bandwidth memory paired with GPUs - we will continue to see new ways to cache and do smarter offloading to solve this

and I do have a port co working on this 😄

#yes AI replaces consultants versus software fully

#from Goldman Sachs IT survey - who pays for AI inference? Net new or taken away from other budgets?

#this is the future but so early

Markets

#Zirpicorns 🦄! AI embedded and accelerating - LFG

First huge congrats to What’s 🔥 reader Jack Naglieri of Panther, a leading AI SOC platform, which just sold to Databricks for $1.4B years after its last funding at a $1.4B valuation back in December 2021. You may say that is not an impressive valuation for a company valued at that price 5 years ago, but I would say it’s an insane exit to be able to grow into the ZIRP-era valuation and be relevant for today’s agentic world!

Congrats to Intercom as well - the poster child for how a founder comes back as CEO, goes founder mode, makes a big bet a few months after ChatGPT launch and 3 years later sold for $3.6B!

#but for the agent native pure startups, the goalposts just keep getting bigger

#the Elon premium 🚀

#with AI products and without

#👀

#🤔

Sometimes the light's all shinin' on me
Other times I can barely see
Lately it occurs to me
What a long, strange trip it's been

It’s truly an insane time as we have the largest IPO in history with SpaceX now worth more than $2 trillion, and the largest private company in the world, Anthropic, forced to take its newest models offline after the U.S. government restricted access. Meanwhile, the market is moving so fast and the uncertainty so high, software companies that would have been funded 12 months ago are having a hard time raising capital.

This is the startup market in June 2026.

At the top, the outcomes have never been bigger. Capital is flooding into frontier AI, chips, data centers, robotics, defense, energy, and space. For a handful of companies, the light has never shined brighter.

Earlier this year, one of our boldstart LPs, StepStone, shared this chart at its annual meeting. It captures just how concentrated venture has become: the top 10 private venture-backed companies represented 112% of the value of all venture exits over the prior decade.

And this was before SpaceX’s IPO and Anthropic’s recent rounds.

In fact, here is what it looks like now - thanks Fable.

Just two companies now represent more value than every venture-backed exit from the last decade combined 🤯. The concentration is simply staggering.

Hardware is simply easier for investors to underwrite right now. It’s physical and scarce, and building it requires enormous capital and deep technical expertise. For many VCs, software is much harder to underwrite but IMO that’s where the opportunity lies!

Every investor I talk to is asking some version of the same questions: What happens when the next model drops? Does the model provider absorb your product? Does your edge disappear overnight? If everyone can build software faster and cheaper, where does lasting value accrue?

It’s easier for many investors to do nothing and just wait until they see signs of a breakout or more of an inflection point. This means many companies with real customers and real revenue are struggling to raise because nobody can say with certainty what software will be worth in three years.

But I believe that uncertainty is exactly where some of the biggest opportunities will be built.

Look at what happened with Anthropic and Fable. Before government restrictions disrupted access, enterprises were already wrestling with rising token costs and limited control over how models could be used. Even Meta is reportedly moving to curb employee token usage as its internal AI costs climb. This is exactly why enterprises cannot rely on a single model provider.

As I’ve said repeatedly before (What’s 🔥 #500 and 501) the opportunity is in applications that work across multiple models, route workloads intelligently, protect proprietary data, manage costs, preserve institutional context, and keep running when a provider or policy changes underneath them. This is only going to accelerate with this latest government intervention - enterprises and founders need more control over their models and data.

Satya Nadella recently offered a framework that I think nails where the value is heading. He argued that a company’s private evals may ultimately become its most valuable IP. (h/t to Gokul for summarizing this)

What’s an eval? Think of it as your company’s custom exam for AI. Not a generic benchmark, but your specific test: Did the model approve the right refund, catch the compliance violation, prioritize the correct security alert, and complete the job at a cost that makes sense? Generic benchmarks tell you which model is smartest. A private eval tells you which model is best for your business. Satya’s test for control is simple: run your eval on model A, swap in model B, and see whether you can keep improving. If you can, you own the application. If you can’t, the model owns you.

Models may become interchangeable sources of raw intelligence, but your private eval captures something they don’t have: your proprietary definition of good, built from years of customer feedback, failures, edge cases, regulatory requirements, and expert judgment. The model supplies the intelligence; the eval defines what valuable intelligence means. And the eval isn’t the whole moat. The real defensibility comes from private evals combined with proprietary context, workflow integration, customer feedback, and execution data. Every completed workflow improves the eval, which improves the product, drives more usage, and generates more feedback. That compounding flywheel is a capability every AI-native startup, SaaS company, and enterprise will need.

This is also why I keep coming back to ROI per workflow as the metric that matters.

The eval measures whether the workflow works, while the software determines which model delivers the best outcome at the right price.

Hardware may be easier to underwrite because its barriers are visible. But software will own the context, evaluation, orchestration, and workflows that turn intelligence into enterprise value. That’s a massive opportunity, and much of the market is still too uncertain to see it.

So yes, this is one of the hardest times I can remember to raise money for a software company. But for the right software companies, hard to underwrite doesn’t mean the opportunity won’t be massive.

What a long, strange trip it’s been.

And somewhere inside all this uncertainty, enormous software companies are being built.

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#love this…

#and this - Elon himself gave SpaceX less than a 10% chance of succeeding:

“I gave SpaceX less than a 10% chance of succeeding at all, to be clear. In fact, I told people this: I said, ‘look, we’re probably going to fail, but, you know, we should give it a try because if we don’t, if if there’s not a new company that enters space, we will never be a truly spacefaring civilization.’”

#there’s another one like this out there now!

Enterprise Tech

#🤯 Claude Fable 5 is out or Mythos for everyone but while super impressive, Anthropic also partially botched the release. First here’s what’s amazing: Felix talks about the third era of AI: I believe we're about to see a major shift, moving from giving AI tasks to giving it responsibilities.

#no longer…must read as the government now has shut down access to foreign countries and foreign nationals, given Anthropic can’t easily identify those it shut down pure access - i suspect this will be resolved by next week but the implications are massive and will only accelerate the idea of control your AI and data

and the view from other countries like India - here’s Zoho’s founder - open source = more control

#💯 before the government intervention, the Fable release and how it was done was already going to accelerate the need for more control of your AI

#btw, you don’t want to get on the Anthropic enterprise tier as its all token consumption

#🤔

#what OpenAI is thinking about cybersecurity with their new hire from Semgrep

#need more original data! Model collapse: AIs trained on AI outputs lose rare, creative “tails” of data first. Generations converge to bland, homogenized slop which is irreversible even with some human data mixed in.

#your data, your AI - Meta also moving from open source to closed and focusing on personal AI

#speaking of personal AI, not sure if I want Meta reading all my data though - look at this chart from Clem below, massive improvements!

#Brian from Coinbase nails it

#🎯 robotic model performance about to make some major step changes - watch Pete Florence from Generalist AI below

#in person experiences as the hedge against AI? It’s why the price of sports teams and other physical assets are going (FT)

#great read on Unitree, a word leader in developing robots

Markets

#just wanted to reshare this 🤯

#SaaS not dead for all, just for some

Until now, building AI around your enterprise has been difficult, time consuming, and expensive. The tradeoff was simple: use frontier intelligence and pay for it, or use cheaper models and sacrifice capability.

That tradeoff is starting to break.

The messaging this week was coming from all directions, but it pointed to the same idea: enterprises no longer have to choose between intelligence and economics.

Some are building AI around their own data, workflows, and institutional knowledge. Others are going further, combining proprietary data with open models, post-training, and routing. Different approaches, same destination: more control, better economics, and AI that increasingly reflects the business itself.

A week ago this felt like a pitch. This week it started looking like a roadmap, although we’re still early.

As discussed in last week’s issue #500, the budget pressure is real and it’s forcing the conversation. Listen to Dara at Uber:

Let's look at three vendors to dig in further. First, Palantir. It costs you money, but it’s smarter because it understands your data, workflows, and institutional knowledge while keeping them under your control. Alex Karp lays it out:

Kirkland & Ellis doubles down on that message.

The frontier labs heard it too. Microsoft's answer is to make the model yours. They launched 7 new models, yes they're in the race with OpenAI, Anthropic, and Google, but more importantly read this from Mustafa's post:

With Frontier Tuning, you’re building your own model, trained on your your own data, within your environment, controlled by you. Ok, this is not open source and neither is Palantir but you do get my point - why use a frontier lab model at the highest cost when perhaps i can rearchitect a better way. Microsoft claims that Land-O-Lakes has a 10x more cost efficient model than Open AI’s GPT 5.5.

Here’s the fork worth naming out loud.

“Own your AI” is splitting into two camps.

Context is yours: Palantir and others are helping enterprises build AI around their institutional knowledge, workflows, permissions, and data. The model underneath can change, but the enterprise context becomes the asset.

Model is yours: Nemotron, GLM, and the post-training crowd give you the model, the data, and the weights, with routing on top.

Both beat renting raw frontier intelligence for everything.

Frontier models aren’t going away. But long term I’m betting the biggest winners will combine frontier intelligence with proprietary data, enterprise context, workflow orchestration, and increasingly their own models.

Control compounds.

The company that owns the AI behind its critical enterprise workflows has more control over its economics, its roadmap, and ultimately its competitive advantage.

That’s not a hedge. That’s the call.

Which is why I love Nvidia putting significant dollars behind Nemotron, which is catching up fast and was released under a new MDW open license: open model, open data, and open weights.

Clem from Hugging Face tells it like it is - routing and post-training open-source models gives you smarter, faster, and CHEAPER systems.

We will move to a world like Clem says where post-training with more control of your data and evals will become an important way to build agents. This is also a world we invested behind at boldstart a couple of years ago with Uare.ai. Here’s an excerpt from founder and CEO Rob Locascio talking more about what and why (stay tuned for release soon!)

Everything above is the same argument told in bits. Generalist AI is that argument told in atoms.

Your data, your model, your AI. The catch in the physical world is that there is no Reddit for robotics data. You can’t scrape your way to a foundation model when the foundation doesn’t exist. So the moat isn’t the model, it’s how you manufacture the data no one else has.

Generalist went and built it. 3d printed data hands on real human hands, video of every kind of work getting done, over 500,000 hours of proprietary training data feeding their own model. With GEN-1, deployments are live and the model learns new skills in hours to days. Same playbook as the software camp, just in robots: own the data, own the model, own the outcome.

Which is why I'm 🔥 up that the company announced its most recent round, $400M at a $2B valuation. New lead Radical alongside 8VC, Union Square, Hanabi, Norwest and more, with existing investors Nvidia, boldstart, Spark, Jeff Bezos and more. Angels include Eric Yuan, Bin Lin, Fei-Fei Li, and Naval Ravikant.

We led the inception round in March 2024 with Nvidia because we believed the autonomous enterprise doesn’t stop at software. It extends into the physical world, where the real value of labor lives. GEN-1 is the proof. If you want the founding story, read my partner Ellen Chisa’s post on Pete Florence, Andy Zeng, and Andy Barry, the PhDs, DeepMind, and the lifelong passion behind it.

And as Masa says, the opportunity is simply massive

So back to where we started. What if?

The tradeoff is starting to disappear. Frontier or affordable, owned or rented, smart or cheap. Those were yesterday’s either/ors. This week made them ands.

The question is no longer whether you can own your AI. It’s whether you can afford not to.

The future isn’t frontier versus open. It’s who owns the context, who owns the model, and ultimately who owns the outcome.

As always, 🙏🏼 for reading and please share with your friends and colleagues

Scaling Startups

#follow this thread - so many reposts of horror stories raising from ummm…venture capitalists…

choose your partner wisely, not just the firm, the partner

Enterprise Tech

#let’s all celebrate a major milestone - agentic traffic has finally surpassed human traffic!

#🤯 will AI build itself? Anthropic sees a path to recursive self-improvement - read the data - engineers ship 8× more code/quarter, 80% of codebase is AI-written, coding success at 76%, 52× training speedups, and beats humans 64% on research decisions. This closes the loop for AI to autonomously design better successors. Massive upside…kind of insane!

#don’t count Microsoft out - back and just launched 7 AI models

#i like Jensen’s thinking on this - best time ever to be a software company but make sure it’s built for agents!

#more on that - deep systems of records on most of these or incredible context with the Palantir ontology

#💯 which is why companies need intelligent process mining and mapping to understand workflows in order to figure out how the org works and how to improve

#and also why many cos are using Forward Deployed Engineers (FDEs) to kick off the process as well and more importantly measure what success is - read the full post here

#model optionality increasing

#super cool video and post from Guy Podjarny of Tessl on the Emerging Agentic Stack and why “context and skills are the new code.” Software development is transforming from writing code and implementation to defining intent and instructions. Code is becoming “disposable” and easily regenerated, while context (guiding the AI agent) is becoming the primary unit of work.

along with a deeper discussion with friend James Kaplan, CTO McKinsey, on the future of software development with Guy from Tessl (watch here)

#more of the same but for security harness engineering

#going to be an insane year for robotics - owning the brain, the intelligence layer going to be big business

and Nvidia

#drones are working…

Markets

#the SpaceX IPO roadshow presentation is here!

#speaking of SpaceX - 100x increase in revenue forecasted by 2030!

#SaaS back? just select names which is the correct view - more nuance needed

#that’s some AI concentration

#it’s never as bad as one thinks…or as good - amazing comeback

That’s 500 weeks in a row. So many times I was wiped out, sick, traveling, or just exhausted, but I kept plugging away because writing this newsletter has been invaluable for me personally.

It started simply as notes to myself shared with friends, and over time it evolved into something much more meaningful. I now find it to be a cathartic way to reflect on the week, synthesize my thoughts, connect trends, and think longer term about where the world and technology may be heading.

Thank you all for reading, sharing, and being part of this journey. Here’s to many more issues ahead!

Now back to the regularly scheduled programming…

Ok all, let’s not overreact to every single AI related news announcement. This week it was all about tokenomics and insane costs. As always the reality is more nuanced.

First, here’s what kicked it off…

But people tend to overreact to headlines without digging deeply enough into the underlying context, which is why I agree with Simon here 💯.

Sure, there is so much wasted spend especially when you have token leaderboards versus being focused on hard ROI.

Measuring tokens consumed is measuring effort, not output, and it’s the same mistake enterprises made a decade ago measuring lines of code or hours logged. You get what you measure, and if you measure tokens, you’ll get engineers finding creative ways to burn them.

Here’s the real story. We’re entering Phase 2 of enterprise AI.

Phase 1 was the subsidy era, where frontier labs aggressively absorbed costs to drive adoption and enterprises experimented freely because tokens were effectively subsidized.

Phase 2 is the consumption era, where vendors charge real money, CFOs ask harder questions, and procurement re-enters every renewal discussion.

The winners from here will:

• intelligently route workloads across models

• lean into open source for the 80% of workloads that don’t require frontier intelligence

• reserve expensive inference for workflows that actually move revenue, efficiency, or customer outcomes

Token leaderboards are a Phase 1 artifact.

ROI per workflow is the Phase 2 metric.

This reminds me a bit of how overblown the “Claude Mythos destroys cybersecurity” narrative became earlier this year. Immediate market reactions tend to overshoot reality as Palo Alto Networks, for example, has already bounced back in stock price!

Gary Marcus shares a few more takes from this tokenomics discussion.

The important point many are missing from Gary Marcus’ comments is that he’s not saying agents don’t work. He’s saying enterprises will become more disciplined in how they measure ROI and deploy AI systems. The conversation shifts from token consumption to business outcomes.

And despite all the noise, token consumption itself is still exploding.

The real question is: where does the value accrue?

Regardless of routing strategy, GPU and compute demand still compounds. Look at this chart from Goldman Sachs showing insane token growth in the next few years at 24x!

Sierra recognized this shift early by focusing on outcomes-based pricing rather than token-based pricing.

What’s becoming increasingly clear is that the best enterprise AI companies are already building:

• routing layers

• swappable model infrastructure with an emphasis on open weight

• proprietary orchestration

• hybrid deployment architectures

• on-prem support for sensitive enterprise environments

Some are even deploying open-weight models on-prem for deep customization and control before selectively invoking frontier models when necessary. Harvey, for example, has used its insane growth to build a powerful proprietary data flywheel and is showing how this gets done.

This is exactly where proprietary enterprise context becomes incredibly valuable.

As open-weight models improve, we’ll increasingly see more and more software vendors and enterprises train and customize models directly on private internal data, routing across a mixture of open and frontier systems depending on sensitivity, latency, and economics.

That’s exactly what Larry Ellison has been talking about.

LFG - it’s still early!

As always, 🙏🏼 for reading and please share with your friends and colleagues!

Scaling Startups

#🎯

#founder advice I strongly concur with

#fun time, lots of parallels between early days of NY and Miami now except timeline is significantly accelerated - listen in or watch

#physical AI

#got to admire the all-in nature of Masa!

Enterprise Tech

#the skeptical view of agents and AI has been all the rage this week…

#remember Facebook for Work, its enterprise play - well that didn’t work out but now its back with its own FDEs but who would want theirs?

#as I’ve written before - What’s 🔥 #498, larger companies are not only using a constellation of models (SOTA, open source) for token costs but also for creating differentiation - why give my employees brains to a model when I can have my own - this will be a bigger and bigger opportunity over time and also may offset billable hours to outcome based pricing?

#seems likes its working…

#🎯

#🤯 insane growth for inference providers who don’t own their own GPUs - thinner margins but slick software on top to optimize inference, etc - all of them growing 📈

#no surprise Fireworks on this list - great list (congrats Keycard, a boldstart port co) and solid deck in this thread - worth a read

#edge inference coming in a big way - Apple will finally flex its muscles with announcments at its WWDC (The Information)

I spoke about this a couple years ago on this pod!

#I feel this every day

#why we led the inception round for Generalist AI 2 years ago and investing more in the stack - stay tuned 👀

#no thank you - but yes, there is no reddit for robotics data

#not a surprise but consulting is going to rapidly change

#🤣

Markets

#reminder - software not dead yet - market overreacted to every SaaS co from frontier lab fear but many have bounced back

#from - Palo Alto back at 15.9x forward with 29% NTM revenue growth and Datadog at 16.9x with 23% revenue growth